ChebMoE: A Spectral-Aware and Expert-Adaptive Framework for Graph Anomaly Detection

ICLR 2026 Conference Submission18638 Authors

Keywords: Anomaly Detection

Abstract:

Graph anomaly detection is critical for applications such as social networks, cybersecurity, and finance, yet remains challenging due to the unique spectral signatures of anomalies. In particular, anomalous nodes often exhibit high-frequency spectral patterns—a phenomenon known as \textit{spectral shift}—which are easily suppressed by the low-pass nature of standard Graph Neural Networks (GNNs), resulting in \textit{spectral washing} and poor anomaly detection.
In this work, we present ChebMoE, a novel and principled framework that directly addresses the spectral limitations of existing GNN-based anomaly detectors. Our key contributions are as follows:
(1) We introduce a Chebyshev polynomial-based spectral feature extractor that efficiently preserves and amplifies high-frequency components, enabling the model to capture subtle spectral shifts associated with anomalies without requiring costly eigendecomposition.
(2) We design a Mixture of Experts (MoE) anomaly detector with a learnable gating mechanism, allowing the model to adaptively aggregate diverse expert subnetworks and flexibly model complex anomaly patterns.
(3) We propose a contrastive anomaly feature generator that leverages self-supervised contrastive learning to further enhance the discriminative power of node representations, improving robustness in the absence of labeled anomalies.
Extensive experiments on seven real-world dynamic graph datasets demonstrate that ChebMoE consistently outperforms state-of-the-art baselines. For example, it achieves ROC-AUC of 0.9906 on Wiki, 0.8791 on Reddit, and 0.9812 on UCI, along with consistently high F1-scores, effectively counteracting spectral washing and substantially advancing the state of graph anomaly detection.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 18638